xref: /freebsd/sys/contrib/openzfs/module/os/linux/spl/spl-taskq.c (revision 6ba2210ee039f2f12878c217bcf058e9c8b26b29)
1 /*
2  *  Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
3  *  Copyright (C) 2007 The Regents of the University of California.
4  *  Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
5  *  Written by Brian Behlendorf <behlendorf1@llnl.gov>.
6  *  UCRL-CODE-235197
7  *
8  *  This file is part of the SPL, Solaris Porting Layer.
9  *
10  *  The SPL is free software; you can redistribute it and/or modify it
11  *  under the terms of the GNU General Public License as published by the
12  *  Free Software Foundation; either version 2 of the License, or (at your
13  *  option) any later version.
14  *
15  *  The SPL is distributed in the hope that it will be useful, but WITHOUT
16  *  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17  *  FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
18  *  for more details.
19  *
20  *  You should have received a copy of the GNU General Public License along
21  *  with the SPL.  If not, see <http://www.gnu.org/licenses/>.
22  *
23  *  Solaris Porting Layer (SPL) Task Queue Implementation.
24  */
25 
26 #include <sys/timer.h>
27 #include <sys/taskq.h>
28 #include <sys/kmem.h>
29 #include <sys/tsd.h>
30 #include <sys/trace_spl.h>
31 #ifdef HAVE_CPU_HOTPLUG
32 #include <linux/cpuhotplug.h>
33 #endif
34 
35 int spl_taskq_thread_bind = 0;
36 module_param(spl_taskq_thread_bind, int, 0644);
37 MODULE_PARM_DESC(spl_taskq_thread_bind, "Bind taskq thread to CPU by default");
38 
39 
40 int spl_taskq_thread_dynamic = 1;
41 module_param(spl_taskq_thread_dynamic, int, 0444);
42 MODULE_PARM_DESC(spl_taskq_thread_dynamic, "Allow dynamic taskq threads");
43 
44 int spl_taskq_thread_priority = 1;
45 module_param(spl_taskq_thread_priority, int, 0644);
46 MODULE_PARM_DESC(spl_taskq_thread_priority,
47 	"Allow non-default priority for taskq threads");
48 
49 int spl_taskq_thread_sequential = 4;
50 module_param(spl_taskq_thread_sequential, int, 0644);
51 MODULE_PARM_DESC(spl_taskq_thread_sequential,
52 	"Create new taskq threads after N sequential tasks");
53 
54 /* Global system-wide dynamic task queue available for all consumers */
55 taskq_t *system_taskq;
56 EXPORT_SYMBOL(system_taskq);
57 /* Global dynamic task queue for long delay */
58 taskq_t *system_delay_taskq;
59 EXPORT_SYMBOL(system_delay_taskq);
60 
61 /* Private dedicated taskq for creating new taskq threads on demand. */
62 static taskq_t *dynamic_taskq;
63 static taskq_thread_t *taskq_thread_create(taskq_t *);
64 
65 #ifdef HAVE_CPU_HOTPLUG
66 /* Multi-callback id for cpu hotplugging. */
67 static int spl_taskq_cpuhp_state;
68 #endif
69 
70 /* List of all taskqs */
71 LIST_HEAD(tq_list);
72 struct rw_semaphore tq_list_sem;
73 static uint_t taskq_tsd;
74 
75 static int
76 task_km_flags(uint_t flags)
77 {
78 	if (flags & TQ_NOSLEEP)
79 		return (KM_NOSLEEP);
80 
81 	if (flags & TQ_PUSHPAGE)
82 		return (KM_PUSHPAGE);
83 
84 	return (KM_SLEEP);
85 }
86 
87 /*
88  * taskq_find_by_name - Find the largest instance number of a named taskq.
89  */
90 static int
91 taskq_find_by_name(const char *name)
92 {
93 	struct list_head *tql = NULL;
94 	taskq_t *tq;
95 
96 	list_for_each_prev(tql, &tq_list) {
97 		tq = list_entry(tql, taskq_t, tq_taskqs);
98 		if (strcmp(name, tq->tq_name) == 0)
99 			return (tq->tq_instance);
100 	}
101 	return (-1);
102 }
103 
104 /*
105  * NOTE: Must be called with tq->tq_lock held, returns a list_t which
106  * is not attached to the free, work, or pending taskq lists.
107  */
108 static taskq_ent_t *
109 task_alloc(taskq_t *tq, uint_t flags, unsigned long *irqflags)
110 {
111 	taskq_ent_t *t;
112 	int count = 0;
113 
114 	ASSERT(tq);
115 retry:
116 	/* Acquire taskq_ent_t's from free list if available */
117 	if (!list_empty(&tq->tq_free_list) && !(flags & TQ_NEW)) {
118 		t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
119 
120 		ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
121 		ASSERT(!(t->tqent_flags & TQENT_FLAG_CANCEL));
122 		ASSERT(!timer_pending(&t->tqent_timer));
123 
124 		list_del_init(&t->tqent_list);
125 		return (t);
126 	}
127 
128 	/* Free list is empty and memory allocations are prohibited */
129 	if (flags & TQ_NOALLOC)
130 		return (NULL);
131 
132 	/* Hit maximum taskq_ent_t pool size */
133 	if (tq->tq_nalloc >= tq->tq_maxalloc) {
134 		if (flags & TQ_NOSLEEP)
135 			return (NULL);
136 
137 		/*
138 		 * Sleep periodically polling the free list for an available
139 		 * taskq_ent_t. Dispatching with TQ_SLEEP should always succeed
140 		 * but we cannot block forever waiting for an taskq_ent_t to
141 		 * show up in the free list, otherwise a deadlock can happen.
142 		 *
143 		 * Therefore, we need to allocate a new task even if the number
144 		 * of allocated tasks is above tq->tq_maxalloc, but we still
145 		 * end up delaying the task allocation by one second, thereby
146 		 * throttling the task dispatch rate.
147 		 */
148 		spin_unlock_irqrestore(&tq->tq_lock, *irqflags);
149 		schedule_timeout(HZ / 100);
150 		spin_lock_irqsave_nested(&tq->tq_lock, *irqflags,
151 		    tq->tq_lock_class);
152 		if (count < 100) {
153 			count++;
154 			goto retry;
155 		}
156 	}
157 
158 	spin_unlock_irqrestore(&tq->tq_lock, *irqflags);
159 	t = kmem_alloc(sizeof (taskq_ent_t), task_km_flags(flags));
160 	spin_lock_irqsave_nested(&tq->tq_lock, *irqflags, tq->tq_lock_class);
161 
162 	if (t) {
163 		taskq_init_ent(t);
164 		tq->tq_nalloc++;
165 	}
166 
167 	return (t);
168 }
169 
170 /*
171  * NOTE: Must be called with tq->tq_lock held, expects the taskq_ent_t
172  * to already be removed from the free, work, or pending taskq lists.
173  */
174 static void
175 task_free(taskq_t *tq, taskq_ent_t *t)
176 {
177 	ASSERT(tq);
178 	ASSERT(t);
179 	ASSERT(list_empty(&t->tqent_list));
180 	ASSERT(!timer_pending(&t->tqent_timer));
181 
182 	kmem_free(t, sizeof (taskq_ent_t));
183 	tq->tq_nalloc--;
184 }
185 
186 /*
187  * NOTE: Must be called with tq->tq_lock held, either destroys the
188  * taskq_ent_t if too many exist or moves it to the free list for later use.
189  */
190 static void
191 task_done(taskq_t *tq, taskq_ent_t *t)
192 {
193 	ASSERT(tq);
194 	ASSERT(t);
195 
196 	/* Wake tasks blocked in taskq_wait_id() */
197 	wake_up_all(&t->tqent_waitq);
198 
199 	list_del_init(&t->tqent_list);
200 
201 	if (tq->tq_nalloc <= tq->tq_minalloc) {
202 		t->tqent_id = TASKQID_INVALID;
203 		t->tqent_func = NULL;
204 		t->tqent_arg = NULL;
205 		t->tqent_flags = 0;
206 
207 		list_add_tail(&t->tqent_list, &tq->tq_free_list);
208 	} else {
209 		task_free(tq, t);
210 	}
211 }
212 
213 /*
214  * When a delayed task timer expires remove it from the delay list and
215  * add it to the priority list in order for immediate processing.
216  */
217 static void
218 task_expire_impl(taskq_ent_t *t)
219 {
220 	taskq_ent_t *w;
221 	taskq_t *tq = t->tqent_taskq;
222 	struct list_head *l = NULL;
223 	unsigned long flags;
224 
225 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
226 
227 	if (t->tqent_flags & TQENT_FLAG_CANCEL) {
228 		ASSERT(list_empty(&t->tqent_list));
229 		spin_unlock_irqrestore(&tq->tq_lock, flags);
230 		return;
231 	}
232 
233 	t->tqent_birth = jiffies;
234 	DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
235 
236 	/*
237 	 * The priority list must be maintained in strict task id order
238 	 * from lowest to highest for lowest_id to be easily calculable.
239 	 */
240 	list_del(&t->tqent_list);
241 	list_for_each_prev(l, &tq->tq_prio_list) {
242 		w = list_entry(l, taskq_ent_t, tqent_list);
243 		if (w->tqent_id < t->tqent_id) {
244 			list_add(&t->tqent_list, l);
245 			break;
246 		}
247 	}
248 	if (l == &tq->tq_prio_list)
249 		list_add(&t->tqent_list, &tq->tq_prio_list);
250 
251 	spin_unlock_irqrestore(&tq->tq_lock, flags);
252 
253 	wake_up(&tq->tq_work_waitq);
254 }
255 
256 static void
257 task_expire(spl_timer_list_t tl)
258 {
259 	struct timer_list *tmr = (struct timer_list *)tl;
260 	taskq_ent_t *t = from_timer(t, tmr, tqent_timer);
261 	task_expire_impl(t);
262 }
263 
264 /*
265  * Returns the lowest incomplete taskqid_t.  The taskqid_t may
266  * be queued on the pending list, on the priority list, on the
267  * delay list, or on the work list currently being handled, but
268  * it is not 100% complete yet.
269  */
270 static taskqid_t
271 taskq_lowest_id(taskq_t *tq)
272 {
273 	taskqid_t lowest_id = tq->tq_next_id;
274 	taskq_ent_t *t;
275 	taskq_thread_t *tqt;
276 
277 	if (!list_empty(&tq->tq_pend_list)) {
278 		t = list_entry(tq->tq_pend_list.next, taskq_ent_t, tqent_list);
279 		lowest_id = MIN(lowest_id, t->tqent_id);
280 	}
281 
282 	if (!list_empty(&tq->tq_prio_list)) {
283 		t = list_entry(tq->tq_prio_list.next, taskq_ent_t, tqent_list);
284 		lowest_id = MIN(lowest_id, t->tqent_id);
285 	}
286 
287 	if (!list_empty(&tq->tq_delay_list)) {
288 		t = list_entry(tq->tq_delay_list.next, taskq_ent_t, tqent_list);
289 		lowest_id = MIN(lowest_id, t->tqent_id);
290 	}
291 
292 	if (!list_empty(&tq->tq_active_list)) {
293 		tqt = list_entry(tq->tq_active_list.next, taskq_thread_t,
294 		    tqt_active_list);
295 		ASSERT(tqt->tqt_id != TASKQID_INVALID);
296 		lowest_id = MIN(lowest_id, tqt->tqt_id);
297 	}
298 
299 	return (lowest_id);
300 }
301 
302 /*
303  * Insert a task into a list keeping the list sorted by increasing taskqid.
304  */
305 static void
306 taskq_insert_in_order(taskq_t *tq, taskq_thread_t *tqt)
307 {
308 	taskq_thread_t *w;
309 	struct list_head *l = NULL;
310 
311 	ASSERT(tq);
312 	ASSERT(tqt);
313 
314 	list_for_each_prev(l, &tq->tq_active_list) {
315 		w = list_entry(l, taskq_thread_t, tqt_active_list);
316 		if (w->tqt_id < tqt->tqt_id) {
317 			list_add(&tqt->tqt_active_list, l);
318 			break;
319 		}
320 	}
321 	if (l == &tq->tq_active_list)
322 		list_add(&tqt->tqt_active_list, &tq->tq_active_list);
323 }
324 
325 /*
326  * Find and return a task from the given list if it exists.  The list
327  * must be in lowest to highest task id order.
328  */
329 static taskq_ent_t *
330 taskq_find_list(taskq_t *tq, struct list_head *lh, taskqid_t id)
331 {
332 	struct list_head *l = NULL;
333 	taskq_ent_t *t;
334 
335 	list_for_each(l, lh) {
336 		t = list_entry(l, taskq_ent_t, tqent_list);
337 
338 		if (t->tqent_id == id)
339 			return (t);
340 
341 		if (t->tqent_id > id)
342 			break;
343 	}
344 
345 	return (NULL);
346 }
347 
348 /*
349  * Find an already dispatched task given the task id regardless of what
350  * state it is in.  If a task is still pending it will be returned.
351  * If a task is executing, then -EBUSY will be returned instead.
352  * If the task has already been run then NULL is returned.
353  */
354 static taskq_ent_t *
355 taskq_find(taskq_t *tq, taskqid_t id)
356 {
357 	taskq_thread_t *tqt;
358 	struct list_head *l = NULL;
359 	taskq_ent_t *t;
360 
361 	t = taskq_find_list(tq, &tq->tq_delay_list, id);
362 	if (t)
363 		return (t);
364 
365 	t = taskq_find_list(tq, &tq->tq_prio_list, id);
366 	if (t)
367 		return (t);
368 
369 	t = taskq_find_list(tq, &tq->tq_pend_list, id);
370 	if (t)
371 		return (t);
372 
373 	list_for_each(l, &tq->tq_active_list) {
374 		tqt = list_entry(l, taskq_thread_t, tqt_active_list);
375 		if (tqt->tqt_id == id) {
376 			/*
377 			 * Instead of returning tqt_task, we just return a non
378 			 * NULL value to prevent misuse, since tqt_task only
379 			 * has two valid fields.
380 			 */
381 			return (ERR_PTR(-EBUSY));
382 		}
383 	}
384 
385 	return (NULL);
386 }
387 
388 /*
389  * Theory for the taskq_wait_id(), taskq_wait_outstanding(), and
390  * taskq_wait() functions below.
391  *
392  * Taskq waiting is accomplished by tracking the lowest outstanding task
393  * id and the next available task id.  As tasks are dispatched they are
394  * added to the tail of the pending, priority, or delay lists.  As worker
395  * threads become available the tasks are removed from the heads of these
396  * lists and linked to the worker threads.  This ensures the lists are
397  * kept sorted by lowest to highest task id.
398  *
399  * Therefore the lowest outstanding task id can be quickly determined by
400  * checking the head item from all of these lists.  This value is stored
401  * with the taskq as the lowest id.  It only needs to be recalculated when
402  * either the task with the current lowest id completes or is canceled.
403  *
404  * By blocking until the lowest task id exceeds the passed task id the
405  * taskq_wait_outstanding() function can be easily implemented.  Similarly,
406  * by blocking until the lowest task id matches the next task id taskq_wait()
407  * can be implemented.
408  *
409  * Callers should be aware that when there are multiple worked threads it
410  * is possible for larger task ids to complete before smaller ones.  Also
411  * when the taskq contains delay tasks with small task ids callers may
412  * block for a considerable length of time waiting for them to expire and
413  * execute.
414  */
415 static int
416 taskq_wait_id_check(taskq_t *tq, taskqid_t id)
417 {
418 	int rc;
419 	unsigned long flags;
420 
421 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
422 	rc = (taskq_find(tq, id) == NULL);
423 	spin_unlock_irqrestore(&tq->tq_lock, flags);
424 
425 	return (rc);
426 }
427 
428 /*
429  * The taskq_wait_id() function blocks until the passed task id completes.
430  * This does not guarantee that all lower task ids have completed.
431  */
432 void
433 taskq_wait_id(taskq_t *tq, taskqid_t id)
434 {
435 	wait_event(tq->tq_wait_waitq, taskq_wait_id_check(tq, id));
436 }
437 EXPORT_SYMBOL(taskq_wait_id);
438 
439 static int
440 taskq_wait_outstanding_check(taskq_t *tq, taskqid_t id)
441 {
442 	int rc;
443 	unsigned long flags;
444 
445 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
446 	rc = (id < tq->tq_lowest_id);
447 	spin_unlock_irqrestore(&tq->tq_lock, flags);
448 
449 	return (rc);
450 }
451 
452 /*
453  * The taskq_wait_outstanding() function will block until all tasks with a
454  * lower taskqid than the passed 'id' have been completed.  Note that all
455  * task id's are assigned monotonically at dispatch time.  Zero may be
456  * passed for the id to indicate all tasks dispatch up to this point,
457  * but not after, should be waited for.
458  */
459 void
460 taskq_wait_outstanding(taskq_t *tq, taskqid_t id)
461 {
462 	id = id ? id : tq->tq_next_id - 1;
463 	wait_event(tq->tq_wait_waitq, taskq_wait_outstanding_check(tq, id));
464 }
465 EXPORT_SYMBOL(taskq_wait_outstanding);
466 
467 static int
468 taskq_wait_check(taskq_t *tq)
469 {
470 	int rc;
471 	unsigned long flags;
472 
473 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
474 	rc = (tq->tq_lowest_id == tq->tq_next_id);
475 	spin_unlock_irqrestore(&tq->tq_lock, flags);
476 
477 	return (rc);
478 }
479 
480 /*
481  * The taskq_wait() function will block until the taskq is empty.
482  * This means that if a taskq re-dispatches work to itself taskq_wait()
483  * callers will block indefinitely.
484  */
485 void
486 taskq_wait(taskq_t *tq)
487 {
488 	wait_event(tq->tq_wait_waitq, taskq_wait_check(tq));
489 }
490 EXPORT_SYMBOL(taskq_wait);
491 
492 int
493 taskq_member(taskq_t *tq, kthread_t *t)
494 {
495 	return (tq == (taskq_t *)tsd_get_by_thread(taskq_tsd, t));
496 }
497 EXPORT_SYMBOL(taskq_member);
498 
499 taskq_t *
500 taskq_of_curthread(void)
501 {
502 	return (tsd_get(taskq_tsd));
503 }
504 EXPORT_SYMBOL(taskq_of_curthread);
505 
506 /*
507  * Cancel an already dispatched task given the task id.  Still pending tasks
508  * will be immediately canceled, and if the task is active the function will
509  * block until it completes.  Preallocated tasks which are canceled must be
510  * freed by the caller.
511  */
512 int
513 taskq_cancel_id(taskq_t *tq, taskqid_t id)
514 {
515 	taskq_ent_t *t;
516 	int rc = ENOENT;
517 	unsigned long flags;
518 
519 	ASSERT(tq);
520 
521 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
522 	t = taskq_find(tq, id);
523 	if (t && t != ERR_PTR(-EBUSY)) {
524 		list_del_init(&t->tqent_list);
525 		t->tqent_flags |= TQENT_FLAG_CANCEL;
526 
527 		/*
528 		 * When canceling the lowest outstanding task id we
529 		 * must recalculate the new lowest outstanding id.
530 		 */
531 		if (tq->tq_lowest_id == t->tqent_id) {
532 			tq->tq_lowest_id = taskq_lowest_id(tq);
533 			ASSERT3S(tq->tq_lowest_id, >, t->tqent_id);
534 		}
535 
536 		/*
537 		 * The task_expire() function takes the tq->tq_lock so drop
538 		 * drop the lock before synchronously cancelling the timer.
539 		 */
540 		if (timer_pending(&t->tqent_timer)) {
541 			spin_unlock_irqrestore(&tq->tq_lock, flags);
542 			del_timer_sync(&t->tqent_timer);
543 			spin_lock_irqsave_nested(&tq->tq_lock, flags,
544 			    tq->tq_lock_class);
545 		}
546 
547 		if (!(t->tqent_flags & TQENT_FLAG_PREALLOC))
548 			task_done(tq, t);
549 
550 		rc = 0;
551 	}
552 	spin_unlock_irqrestore(&tq->tq_lock, flags);
553 
554 	if (t == ERR_PTR(-EBUSY)) {
555 		taskq_wait_id(tq, id);
556 		rc = EBUSY;
557 	}
558 
559 	return (rc);
560 }
561 EXPORT_SYMBOL(taskq_cancel_id);
562 
563 static int taskq_thread_spawn(taskq_t *tq);
564 
565 taskqid_t
566 taskq_dispatch(taskq_t *tq, task_func_t func, void *arg, uint_t flags)
567 {
568 	taskq_ent_t *t;
569 	taskqid_t rc = TASKQID_INVALID;
570 	unsigned long irqflags;
571 
572 	ASSERT(tq);
573 	ASSERT(func);
574 
575 	spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class);
576 
577 	/* Taskq being destroyed and all tasks drained */
578 	if (!(tq->tq_flags & TASKQ_ACTIVE))
579 		goto out;
580 
581 	/* Do not queue the task unless there is idle thread for it */
582 	ASSERT(tq->tq_nactive <= tq->tq_nthreads);
583 	if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) {
584 		/* Dynamic taskq may be able to spawn another thread */
585 		if (!(tq->tq_flags & TASKQ_DYNAMIC) ||
586 		    taskq_thread_spawn(tq) == 0)
587 			goto out;
588 	}
589 
590 	if ((t = task_alloc(tq, flags, &irqflags)) == NULL)
591 		goto out;
592 
593 	spin_lock(&t->tqent_lock);
594 
595 	/* Queue to the front of the list to enforce TQ_NOQUEUE semantics */
596 	if (flags & TQ_NOQUEUE)
597 		list_add(&t->tqent_list, &tq->tq_prio_list);
598 	/* Queue to the priority list instead of the pending list */
599 	else if (flags & TQ_FRONT)
600 		list_add_tail(&t->tqent_list, &tq->tq_prio_list);
601 	else
602 		list_add_tail(&t->tqent_list, &tq->tq_pend_list);
603 
604 	t->tqent_id = rc = tq->tq_next_id;
605 	tq->tq_next_id++;
606 	t->tqent_func = func;
607 	t->tqent_arg = arg;
608 	t->tqent_taskq = tq;
609 	t->tqent_timer.function = NULL;
610 	t->tqent_timer.expires = 0;
611 
612 	t->tqent_birth = jiffies;
613 	DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
614 
615 	ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
616 
617 	spin_unlock(&t->tqent_lock);
618 
619 	wake_up(&tq->tq_work_waitq);
620 out:
621 	/* Spawn additional taskq threads if required. */
622 	if (!(flags & TQ_NOQUEUE) && tq->tq_nactive == tq->tq_nthreads)
623 		(void) taskq_thread_spawn(tq);
624 
625 	spin_unlock_irqrestore(&tq->tq_lock, irqflags);
626 	return (rc);
627 }
628 EXPORT_SYMBOL(taskq_dispatch);
629 
630 taskqid_t
631 taskq_dispatch_delay(taskq_t *tq, task_func_t func, void *arg,
632     uint_t flags, clock_t expire_time)
633 {
634 	taskqid_t rc = TASKQID_INVALID;
635 	taskq_ent_t *t;
636 	unsigned long irqflags;
637 
638 	ASSERT(tq);
639 	ASSERT(func);
640 
641 	spin_lock_irqsave_nested(&tq->tq_lock, irqflags, tq->tq_lock_class);
642 
643 	/* Taskq being destroyed and all tasks drained */
644 	if (!(tq->tq_flags & TASKQ_ACTIVE))
645 		goto out;
646 
647 	if ((t = task_alloc(tq, flags, &irqflags)) == NULL)
648 		goto out;
649 
650 	spin_lock(&t->tqent_lock);
651 
652 	/* Queue to the delay list for subsequent execution */
653 	list_add_tail(&t->tqent_list, &tq->tq_delay_list);
654 
655 	t->tqent_id = rc = tq->tq_next_id;
656 	tq->tq_next_id++;
657 	t->tqent_func = func;
658 	t->tqent_arg = arg;
659 	t->tqent_taskq = tq;
660 	t->tqent_timer.function = task_expire;
661 	t->tqent_timer.expires = (unsigned long)expire_time;
662 	add_timer(&t->tqent_timer);
663 
664 	ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
665 
666 	spin_unlock(&t->tqent_lock);
667 out:
668 	/* Spawn additional taskq threads if required. */
669 	if (tq->tq_nactive == tq->tq_nthreads)
670 		(void) taskq_thread_spawn(tq);
671 	spin_unlock_irqrestore(&tq->tq_lock, irqflags);
672 	return (rc);
673 }
674 EXPORT_SYMBOL(taskq_dispatch_delay);
675 
676 void
677 taskq_dispatch_ent(taskq_t *tq, task_func_t func, void *arg, uint_t flags,
678     taskq_ent_t *t)
679 {
680 	unsigned long irqflags;
681 	ASSERT(tq);
682 	ASSERT(func);
683 
684 	spin_lock_irqsave_nested(&tq->tq_lock, irqflags,
685 	    tq->tq_lock_class);
686 
687 	/* Taskq being destroyed and all tasks drained */
688 	if (!(tq->tq_flags & TASKQ_ACTIVE)) {
689 		t->tqent_id = TASKQID_INVALID;
690 		goto out;
691 	}
692 
693 	if ((flags & TQ_NOQUEUE) && (tq->tq_nactive == tq->tq_nthreads)) {
694 		/* Dynamic taskq may be able to spawn another thread */
695 		if (!(tq->tq_flags & TASKQ_DYNAMIC) ||
696 		    taskq_thread_spawn(tq) == 0)
697 			goto out2;
698 		flags |= TQ_FRONT;
699 	}
700 
701 	spin_lock(&t->tqent_lock);
702 
703 	/*
704 	 * Make sure the entry is not on some other taskq; it is important to
705 	 * ASSERT() under lock
706 	 */
707 	ASSERT(taskq_empty_ent(t));
708 
709 	/*
710 	 * Mark it as a prealloc'd task.  This is important
711 	 * to ensure that we don't free it later.
712 	 */
713 	t->tqent_flags |= TQENT_FLAG_PREALLOC;
714 
715 	/* Queue to the priority list instead of the pending list */
716 	if (flags & TQ_FRONT)
717 		list_add_tail(&t->tqent_list, &tq->tq_prio_list);
718 	else
719 		list_add_tail(&t->tqent_list, &tq->tq_pend_list);
720 
721 	t->tqent_id = tq->tq_next_id;
722 	tq->tq_next_id++;
723 	t->tqent_func = func;
724 	t->tqent_arg = arg;
725 	t->tqent_taskq = tq;
726 
727 	t->tqent_birth = jiffies;
728 	DTRACE_PROBE1(taskq_ent__birth, taskq_ent_t *, t);
729 
730 	spin_unlock(&t->tqent_lock);
731 
732 	wake_up(&tq->tq_work_waitq);
733 out:
734 	/* Spawn additional taskq threads if required. */
735 	if (tq->tq_nactive == tq->tq_nthreads)
736 		(void) taskq_thread_spawn(tq);
737 out2:
738 	spin_unlock_irqrestore(&tq->tq_lock, irqflags);
739 }
740 EXPORT_SYMBOL(taskq_dispatch_ent);
741 
742 int
743 taskq_empty_ent(taskq_ent_t *t)
744 {
745 	return (list_empty(&t->tqent_list));
746 }
747 EXPORT_SYMBOL(taskq_empty_ent);
748 
749 void
750 taskq_init_ent(taskq_ent_t *t)
751 {
752 	spin_lock_init(&t->tqent_lock);
753 	init_waitqueue_head(&t->tqent_waitq);
754 	timer_setup(&t->tqent_timer, NULL, 0);
755 	INIT_LIST_HEAD(&t->tqent_list);
756 	t->tqent_id = 0;
757 	t->tqent_func = NULL;
758 	t->tqent_arg = NULL;
759 	t->tqent_flags = 0;
760 	t->tqent_taskq = NULL;
761 }
762 EXPORT_SYMBOL(taskq_init_ent);
763 
764 /*
765  * Return the next pending task, preference is given to tasks on the
766  * priority list which were dispatched with TQ_FRONT.
767  */
768 static taskq_ent_t *
769 taskq_next_ent(taskq_t *tq)
770 {
771 	struct list_head *list;
772 
773 	if (!list_empty(&tq->tq_prio_list))
774 		list = &tq->tq_prio_list;
775 	else if (!list_empty(&tq->tq_pend_list))
776 		list = &tq->tq_pend_list;
777 	else
778 		return (NULL);
779 
780 	return (list_entry(list->next, taskq_ent_t, tqent_list));
781 }
782 
783 /*
784  * Spawns a new thread for the specified taskq.
785  */
786 static void
787 taskq_thread_spawn_task(void *arg)
788 {
789 	taskq_t *tq = (taskq_t *)arg;
790 	unsigned long flags;
791 
792 	if (taskq_thread_create(tq) == NULL) {
793 		/* restore spawning count if failed */
794 		spin_lock_irqsave_nested(&tq->tq_lock, flags,
795 		    tq->tq_lock_class);
796 		tq->tq_nspawn--;
797 		spin_unlock_irqrestore(&tq->tq_lock, flags);
798 	}
799 }
800 
801 /*
802  * Spawn addition threads for dynamic taskqs (TASKQ_DYNAMIC) the current
803  * number of threads is insufficient to handle the pending tasks.  These
804  * new threads must be created by the dedicated dynamic_taskq to avoid
805  * deadlocks between thread creation and memory reclaim.  The system_taskq
806  * which is also a dynamic taskq cannot be safely used for this.
807  */
808 static int
809 taskq_thread_spawn(taskq_t *tq)
810 {
811 	int spawning = 0;
812 
813 	if (!(tq->tq_flags & TASKQ_DYNAMIC))
814 		return (0);
815 
816 	if ((tq->tq_nthreads + tq->tq_nspawn < tq->tq_maxthreads) &&
817 	    (tq->tq_flags & TASKQ_ACTIVE)) {
818 		spawning = (++tq->tq_nspawn);
819 		taskq_dispatch(dynamic_taskq, taskq_thread_spawn_task,
820 		    tq, TQ_NOSLEEP);
821 	}
822 
823 	return (spawning);
824 }
825 
826 /*
827  * Threads in a dynamic taskq should only exit once it has been completely
828  * drained and no other threads are actively servicing tasks.  This prevents
829  * threads from being created and destroyed more than is required.
830  *
831  * The first thread is the thread list is treated as the primary thread.
832  * There is nothing special about the primary thread but in order to avoid
833  * all the taskq pids from changing we opt to make it long running.
834  */
835 static int
836 taskq_thread_should_stop(taskq_t *tq, taskq_thread_t *tqt)
837 {
838 	if (!(tq->tq_flags & TASKQ_DYNAMIC))
839 		return (0);
840 
841 	if (list_first_entry(&(tq->tq_thread_list), taskq_thread_t,
842 	    tqt_thread_list) == tqt)
843 		return (0);
844 
845 	return
846 	    ((tq->tq_nspawn == 0) &&	/* No threads are being spawned */
847 	    (tq->tq_nactive == 0) &&	/* No threads are handling tasks */
848 	    (tq->tq_nthreads > 1) &&	/* More than 1 thread is running */
849 	    (!taskq_next_ent(tq)) &&	/* There are no pending tasks */
850 	    (spl_taskq_thread_dynamic)); /* Dynamic taskqs are allowed */
851 }
852 
853 static int
854 taskq_thread(void *args)
855 {
856 	DECLARE_WAITQUEUE(wait, current);
857 	sigset_t blocked;
858 	taskq_thread_t *tqt = args;
859 	taskq_t *tq;
860 	taskq_ent_t *t;
861 	int seq_tasks = 0;
862 	unsigned long flags;
863 	taskq_ent_t dup_task = {};
864 
865 	ASSERT(tqt);
866 	ASSERT(tqt->tqt_tq);
867 	tq = tqt->tqt_tq;
868 	current->flags |= PF_NOFREEZE;
869 
870 	(void) spl_fstrans_mark();
871 
872 	sigfillset(&blocked);
873 	sigprocmask(SIG_BLOCK, &blocked, NULL);
874 	flush_signals(current);
875 
876 	tsd_set(taskq_tsd, tq);
877 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
878 	/*
879 	 * If we are dynamically spawned, decrease spawning count. Note that
880 	 * we could be created during taskq_create, in which case we shouldn't
881 	 * do the decrement. But it's fine because taskq_create will reset
882 	 * tq_nspawn later.
883 	 */
884 	if (tq->tq_flags & TASKQ_DYNAMIC)
885 		tq->tq_nspawn--;
886 
887 	/* Immediately exit if more threads than allowed were created. */
888 	if (tq->tq_nthreads >= tq->tq_maxthreads)
889 		goto error;
890 
891 	tq->tq_nthreads++;
892 	list_add_tail(&tqt->tqt_thread_list, &tq->tq_thread_list);
893 	wake_up(&tq->tq_wait_waitq);
894 	set_current_state(TASK_INTERRUPTIBLE);
895 
896 	while (!kthread_should_stop()) {
897 
898 		if (list_empty(&tq->tq_pend_list) &&
899 		    list_empty(&tq->tq_prio_list)) {
900 
901 			if (taskq_thread_should_stop(tq, tqt)) {
902 				wake_up_all(&tq->tq_wait_waitq);
903 				break;
904 			}
905 
906 			add_wait_queue_exclusive(&tq->tq_work_waitq, &wait);
907 			spin_unlock_irqrestore(&tq->tq_lock, flags);
908 
909 			schedule();
910 			seq_tasks = 0;
911 
912 			spin_lock_irqsave_nested(&tq->tq_lock, flags,
913 			    tq->tq_lock_class);
914 			remove_wait_queue(&tq->tq_work_waitq, &wait);
915 		} else {
916 			__set_current_state(TASK_RUNNING);
917 		}
918 
919 		if ((t = taskq_next_ent(tq)) != NULL) {
920 			list_del_init(&t->tqent_list);
921 
922 			/*
923 			 * A TQENT_FLAG_PREALLOC task may be reused or freed
924 			 * during the task function call. Store tqent_id and
925 			 * tqent_flags here.
926 			 *
927 			 * Also use an on stack taskq_ent_t for tqt_task
928 			 * assignment in this case; we want to make sure
929 			 * to duplicate all fields, so the values are
930 			 * correct when it's accessed via DTRACE_PROBE*.
931 			 */
932 			tqt->tqt_id = t->tqent_id;
933 			tqt->tqt_flags = t->tqent_flags;
934 
935 			if (t->tqent_flags & TQENT_FLAG_PREALLOC) {
936 				dup_task = *t;
937 				t = &dup_task;
938 			}
939 			tqt->tqt_task = t;
940 
941 			taskq_insert_in_order(tq, tqt);
942 			tq->tq_nactive++;
943 			spin_unlock_irqrestore(&tq->tq_lock, flags);
944 
945 			DTRACE_PROBE1(taskq_ent__start, taskq_ent_t *, t);
946 
947 			/* Perform the requested task */
948 			t->tqent_func(t->tqent_arg);
949 
950 			DTRACE_PROBE1(taskq_ent__finish, taskq_ent_t *, t);
951 
952 			spin_lock_irqsave_nested(&tq->tq_lock, flags,
953 			    tq->tq_lock_class);
954 			tq->tq_nactive--;
955 			list_del_init(&tqt->tqt_active_list);
956 			tqt->tqt_task = NULL;
957 
958 			/* For prealloc'd tasks, we don't free anything. */
959 			if (!(tqt->tqt_flags & TQENT_FLAG_PREALLOC))
960 				task_done(tq, t);
961 
962 			/*
963 			 * When the current lowest outstanding taskqid is
964 			 * done calculate the new lowest outstanding id
965 			 */
966 			if (tq->tq_lowest_id == tqt->tqt_id) {
967 				tq->tq_lowest_id = taskq_lowest_id(tq);
968 				ASSERT3S(tq->tq_lowest_id, >, tqt->tqt_id);
969 			}
970 
971 			/* Spawn additional taskq threads if required. */
972 			if ((++seq_tasks) > spl_taskq_thread_sequential &&
973 			    taskq_thread_spawn(tq))
974 				seq_tasks = 0;
975 
976 			tqt->tqt_id = TASKQID_INVALID;
977 			tqt->tqt_flags = 0;
978 			wake_up_all(&tq->tq_wait_waitq);
979 		} else {
980 			if (taskq_thread_should_stop(tq, tqt))
981 				break;
982 		}
983 
984 		set_current_state(TASK_INTERRUPTIBLE);
985 
986 	}
987 
988 	__set_current_state(TASK_RUNNING);
989 	tq->tq_nthreads--;
990 	list_del_init(&tqt->tqt_thread_list);
991 error:
992 	kmem_free(tqt, sizeof (taskq_thread_t));
993 	spin_unlock_irqrestore(&tq->tq_lock, flags);
994 
995 	tsd_set(taskq_tsd, NULL);
996 	thread_exit();
997 
998 	return (0);
999 }
1000 
1001 static taskq_thread_t *
1002 taskq_thread_create(taskq_t *tq)
1003 {
1004 	static int last_used_cpu = 0;
1005 	taskq_thread_t *tqt;
1006 
1007 	tqt = kmem_alloc(sizeof (*tqt), KM_PUSHPAGE);
1008 	INIT_LIST_HEAD(&tqt->tqt_thread_list);
1009 	INIT_LIST_HEAD(&tqt->tqt_active_list);
1010 	tqt->tqt_tq = tq;
1011 	tqt->tqt_id = TASKQID_INVALID;
1012 
1013 	tqt->tqt_thread = spl_kthread_create(taskq_thread, tqt,
1014 	    "%s", tq->tq_name);
1015 	if (tqt->tqt_thread == NULL) {
1016 		kmem_free(tqt, sizeof (taskq_thread_t));
1017 		return (NULL);
1018 	}
1019 
1020 	if (spl_taskq_thread_bind) {
1021 		last_used_cpu = (last_used_cpu + 1) % num_online_cpus();
1022 		kthread_bind(tqt->tqt_thread, last_used_cpu);
1023 	}
1024 
1025 	if (spl_taskq_thread_priority)
1026 		set_user_nice(tqt->tqt_thread, PRIO_TO_NICE(tq->tq_pri));
1027 
1028 	wake_up_process(tqt->tqt_thread);
1029 
1030 	return (tqt);
1031 }
1032 
1033 taskq_t *
1034 taskq_create(const char *name, int threads_arg, pri_t pri,
1035     int minalloc, int maxalloc, uint_t flags)
1036 {
1037 	taskq_t *tq;
1038 	taskq_thread_t *tqt;
1039 	int count = 0, rc = 0, i;
1040 	unsigned long irqflags;
1041 	int nthreads = threads_arg;
1042 
1043 	ASSERT(name != NULL);
1044 	ASSERT(minalloc >= 0);
1045 	ASSERT(maxalloc <= INT_MAX);
1046 	ASSERT(!(flags & (TASKQ_CPR_SAFE))); /* Unsupported */
1047 
1048 	/* Scale the number of threads using nthreads as a percentage */
1049 	if (flags & TASKQ_THREADS_CPU_PCT) {
1050 		ASSERT(nthreads <= 100);
1051 		ASSERT(nthreads >= 0);
1052 		nthreads = MIN(threads_arg, 100);
1053 		nthreads = MAX(nthreads, 0);
1054 		nthreads = MAX((num_online_cpus() * nthreads) /100, 1);
1055 	}
1056 
1057 	tq = kmem_alloc(sizeof (*tq), KM_PUSHPAGE);
1058 	if (tq == NULL)
1059 		return (NULL);
1060 
1061 	tq->tq_hp_support = B_FALSE;
1062 #ifdef HAVE_CPU_HOTPLUG
1063 	if (flags & TASKQ_THREADS_CPU_PCT) {
1064 		tq->tq_hp_support = B_TRUE;
1065 		if (cpuhp_state_add_instance_nocalls(spl_taskq_cpuhp_state,
1066 		    &tq->tq_hp_cb_node) != 0) {
1067 			kmem_free(tq, sizeof (*tq));
1068 			return (NULL);
1069 		}
1070 	}
1071 #endif
1072 
1073 	spin_lock_init(&tq->tq_lock);
1074 	INIT_LIST_HEAD(&tq->tq_thread_list);
1075 	INIT_LIST_HEAD(&tq->tq_active_list);
1076 	tq->tq_name = kmem_strdup(name);
1077 	tq->tq_nactive = 0;
1078 	tq->tq_nthreads = 0;
1079 	tq->tq_nspawn = 0;
1080 	tq->tq_maxthreads = nthreads;
1081 	tq->tq_cpu_pct = threads_arg;
1082 	tq->tq_pri = pri;
1083 	tq->tq_minalloc = minalloc;
1084 	tq->tq_maxalloc = maxalloc;
1085 	tq->tq_nalloc = 0;
1086 	tq->tq_flags = (flags | TASKQ_ACTIVE);
1087 	tq->tq_next_id = TASKQID_INITIAL;
1088 	tq->tq_lowest_id = TASKQID_INITIAL;
1089 	INIT_LIST_HEAD(&tq->tq_free_list);
1090 	INIT_LIST_HEAD(&tq->tq_pend_list);
1091 	INIT_LIST_HEAD(&tq->tq_prio_list);
1092 	INIT_LIST_HEAD(&tq->tq_delay_list);
1093 	init_waitqueue_head(&tq->tq_work_waitq);
1094 	init_waitqueue_head(&tq->tq_wait_waitq);
1095 	tq->tq_lock_class = TQ_LOCK_GENERAL;
1096 	INIT_LIST_HEAD(&tq->tq_taskqs);
1097 
1098 	if (flags & TASKQ_PREPOPULATE) {
1099 		spin_lock_irqsave_nested(&tq->tq_lock, irqflags,
1100 		    tq->tq_lock_class);
1101 
1102 		for (i = 0; i < minalloc; i++)
1103 			task_done(tq, task_alloc(tq, TQ_PUSHPAGE | TQ_NEW,
1104 			    &irqflags));
1105 
1106 		spin_unlock_irqrestore(&tq->tq_lock, irqflags);
1107 	}
1108 
1109 	if ((flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic)
1110 		nthreads = 1;
1111 
1112 	for (i = 0; i < nthreads; i++) {
1113 		tqt = taskq_thread_create(tq);
1114 		if (tqt == NULL)
1115 			rc = 1;
1116 		else
1117 			count++;
1118 	}
1119 
1120 	/* Wait for all threads to be started before potential destroy */
1121 	wait_event(tq->tq_wait_waitq, tq->tq_nthreads == count);
1122 	/*
1123 	 * taskq_thread might have touched nspawn, but we don't want them to
1124 	 * because they're not dynamically spawned. So we reset it to 0
1125 	 */
1126 	tq->tq_nspawn = 0;
1127 
1128 	if (rc) {
1129 		taskq_destroy(tq);
1130 		tq = NULL;
1131 	} else {
1132 		down_write(&tq_list_sem);
1133 		tq->tq_instance = taskq_find_by_name(name) + 1;
1134 		list_add_tail(&tq->tq_taskqs, &tq_list);
1135 		up_write(&tq_list_sem);
1136 	}
1137 
1138 	return (tq);
1139 }
1140 EXPORT_SYMBOL(taskq_create);
1141 
1142 void
1143 taskq_destroy(taskq_t *tq)
1144 {
1145 	struct task_struct *thread;
1146 	taskq_thread_t *tqt;
1147 	taskq_ent_t *t;
1148 	unsigned long flags;
1149 
1150 	ASSERT(tq);
1151 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1152 	tq->tq_flags &= ~TASKQ_ACTIVE;
1153 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1154 
1155 #ifdef HAVE_CPU_HOTPLUG
1156 	if (tq->tq_hp_support) {
1157 		VERIFY0(cpuhp_state_remove_instance_nocalls(
1158 		    spl_taskq_cpuhp_state, &tq->tq_hp_cb_node));
1159 	}
1160 #endif
1161 	/*
1162 	 * When TASKQ_ACTIVE is clear new tasks may not be added nor may
1163 	 * new worker threads be spawned for dynamic taskq.
1164 	 */
1165 	if (dynamic_taskq != NULL)
1166 		taskq_wait_outstanding(dynamic_taskq, 0);
1167 
1168 	taskq_wait(tq);
1169 
1170 	/* remove taskq from global list used by the kstats */
1171 	down_write(&tq_list_sem);
1172 	list_del(&tq->tq_taskqs);
1173 	up_write(&tq_list_sem);
1174 
1175 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1176 	/* wait for spawning threads to insert themselves to the list */
1177 	while (tq->tq_nspawn) {
1178 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1179 		schedule_timeout_interruptible(1);
1180 		spin_lock_irqsave_nested(&tq->tq_lock, flags,
1181 		    tq->tq_lock_class);
1182 	}
1183 
1184 	/*
1185 	 * Signal each thread to exit and block until it does.  Each thread
1186 	 * is responsible for removing itself from the list and freeing its
1187 	 * taskq_thread_t.  This allows for idle threads to opt to remove
1188 	 * themselves from the taskq.  They can be recreated as needed.
1189 	 */
1190 	while (!list_empty(&tq->tq_thread_list)) {
1191 		tqt = list_entry(tq->tq_thread_list.next,
1192 		    taskq_thread_t, tqt_thread_list);
1193 		thread = tqt->tqt_thread;
1194 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1195 
1196 		kthread_stop(thread);
1197 
1198 		spin_lock_irqsave_nested(&tq->tq_lock, flags,
1199 		    tq->tq_lock_class);
1200 	}
1201 
1202 	while (!list_empty(&tq->tq_free_list)) {
1203 		t = list_entry(tq->tq_free_list.next, taskq_ent_t, tqent_list);
1204 
1205 		ASSERT(!(t->tqent_flags & TQENT_FLAG_PREALLOC));
1206 
1207 		list_del_init(&t->tqent_list);
1208 		task_free(tq, t);
1209 	}
1210 
1211 	ASSERT0(tq->tq_nthreads);
1212 	ASSERT0(tq->tq_nalloc);
1213 	ASSERT0(tq->tq_nspawn);
1214 	ASSERT(list_empty(&tq->tq_thread_list));
1215 	ASSERT(list_empty(&tq->tq_active_list));
1216 	ASSERT(list_empty(&tq->tq_free_list));
1217 	ASSERT(list_empty(&tq->tq_pend_list));
1218 	ASSERT(list_empty(&tq->tq_prio_list));
1219 	ASSERT(list_empty(&tq->tq_delay_list));
1220 
1221 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1222 
1223 	kmem_strfree(tq->tq_name);
1224 	kmem_free(tq, sizeof (taskq_t));
1225 }
1226 EXPORT_SYMBOL(taskq_destroy);
1227 
1228 static unsigned int spl_taskq_kick = 0;
1229 
1230 /*
1231  * 2.6.36 API Change
1232  * module_param_cb is introduced to take kernel_param_ops and
1233  * module_param_call is marked as obsolete. Also set and get operations
1234  * were changed to take a 'const struct kernel_param *'.
1235  */
1236 static int
1237 #ifdef module_param_cb
1238 param_set_taskq_kick(const char *val, const struct kernel_param *kp)
1239 #else
1240 param_set_taskq_kick(const char *val, struct kernel_param *kp)
1241 #endif
1242 {
1243 	int ret;
1244 	taskq_t *tq = NULL;
1245 	taskq_ent_t *t;
1246 	unsigned long flags;
1247 
1248 	ret = param_set_uint(val, kp);
1249 	if (ret < 0 || !spl_taskq_kick)
1250 		return (ret);
1251 	/* reset value */
1252 	spl_taskq_kick = 0;
1253 
1254 	down_read(&tq_list_sem);
1255 	list_for_each_entry(tq, &tq_list, tq_taskqs) {
1256 		spin_lock_irqsave_nested(&tq->tq_lock, flags,
1257 		    tq->tq_lock_class);
1258 		/* Check if the first pending is older than 5 seconds */
1259 		t = taskq_next_ent(tq);
1260 		if (t && time_after(jiffies, t->tqent_birth + 5*HZ)) {
1261 			(void) taskq_thread_spawn(tq);
1262 			printk(KERN_INFO "spl: Kicked taskq %s/%d\n",
1263 			    tq->tq_name, tq->tq_instance);
1264 		}
1265 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1266 	}
1267 	up_read(&tq_list_sem);
1268 	return (ret);
1269 }
1270 
1271 #ifdef module_param_cb
1272 static const struct kernel_param_ops param_ops_taskq_kick = {
1273 	.set = param_set_taskq_kick,
1274 	.get = param_get_uint,
1275 };
1276 module_param_cb(spl_taskq_kick, &param_ops_taskq_kick, &spl_taskq_kick, 0644);
1277 #else
1278 module_param_call(spl_taskq_kick, param_set_taskq_kick, param_get_uint,
1279 	&spl_taskq_kick, 0644);
1280 #endif
1281 MODULE_PARM_DESC(spl_taskq_kick,
1282 	"Write nonzero to kick stuck taskqs to spawn more threads");
1283 
1284 #ifdef HAVE_CPU_HOTPLUG
1285 /*
1286  * This callback will be called exactly once for each core that comes online,
1287  * for each dynamic taskq. We attempt to expand taskqs that have
1288  * TASKQ_THREADS_CPU_PCT set. We need to redo the percentage calculation every
1289  * time, to correctly determine whether or not to add a thread.
1290  */
1291 static int
1292 spl_taskq_expand(unsigned int cpu, struct hlist_node *node)
1293 {
1294 	taskq_t *tq = list_entry(node, taskq_t, tq_hp_cb_node);
1295 	unsigned long flags;
1296 	int err = 0;
1297 
1298 	ASSERT(tq);
1299 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1300 
1301 	if (!(tq->tq_flags & TASKQ_ACTIVE))
1302 		goto out;
1303 
1304 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
1305 	int nthreads = MIN(tq->tq_cpu_pct, 100);
1306 	nthreads = MAX(((num_online_cpus() + 1) * nthreads) / 100, 1);
1307 	tq->tq_maxthreads = nthreads;
1308 
1309 	if (!((tq->tq_flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) &&
1310 	    tq->tq_maxthreads > tq->tq_nthreads) {
1311 		ASSERT3U(tq->tq_maxthreads, ==, tq->tq_nthreads + 1);
1312 		taskq_thread_t *tqt = taskq_thread_create(tq);
1313 		if (tqt == NULL)
1314 			err = -1;
1315 	}
1316 
1317 out:
1318 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1319 	return (err);
1320 }
1321 
1322 /*
1323  * While we don't support offlining CPUs, it is possible that CPUs will fail
1324  * to online successfully. We do need to be able to handle this case
1325  * gracefully.
1326  */
1327 static int
1328 spl_taskq_prepare_down(unsigned int cpu, struct hlist_node *node)
1329 {
1330 	taskq_t *tq = list_entry(node, taskq_t, tq_hp_cb_node);
1331 	unsigned long flags;
1332 
1333 	ASSERT(tq);
1334 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);
1335 
1336 	if (!(tq->tq_flags & TASKQ_ACTIVE))
1337 		goto out;
1338 
1339 	ASSERT(tq->tq_flags & TASKQ_THREADS_CPU_PCT);
1340 	int nthreads = MIN(tq->tq_cpu_pct, 100);
1341 	nthreads = MAX(((num_online_cpus()) * nthreads) / 100, 1);
1342 	tq->tq_maxthreads = nthreads;
1343 
1344 	if (!((tq->tq_flags & TASKQ_DYNAMIC) && spl_taskq_thread_dynamic) &&
1345 	    tq->tq_maxthreads < tq->tq_nthreads) {
1346 		ASSERT3U(tq->tq_maxthreads, ==, tq->tq_nthreads - 1);
1347 		taskq_thread_t *tqt = list_entry(tq->tq_thread_list.next,
1348 		    taskq_thread_t, tqt_thread_list);
1349 		struct task_struct *thread = tqt->tqt_thread;
1350 		spin_unlock_irqrestore(&tq->tq_lock, flags);
1351 
1352 		kthread_stop(thread);
1353 
1354 		return (0);
1355 	}
1356 
1357 out:
1358 	spin_unlock_irqrestore(&tq->tq_lock, flags);
1359 	return (0);
1360 }
1361 #endif
1362 
1363 int
1364 spl_taskq_init(void)
1365 {
1366 	init_rwsem(&tq_list_sem);
1367 	tsd_create(&taskq_tsd, NULL);
1368 
1369 #ifdef HAVE_CPU_HOTPLUG
1370 	spl_taskq_cpuhp_state = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1371 	    "fs/spl_taskq:online", spl_taskq_expand, spl_taskq_prepare_down);
1372 #endif
1373 
1374 	system_taskq = taskq_create("spl_system_taskq", MAX(boot_ncpus, 64),
1375 	    maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC);
1376 	if (system_taskq == NULL)
1377 		return (1);
1378 
1379 	system_delay_taskq = taskq_create("spl_delay_taskq", MAX(boot_ncpus, 4),
1380 	    maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE|TASKQ_DYNAMIC);
1381 	if (system_delay_taskq == NULL) {
1382 #ifdef HAVE_CPU_HOTPLUG
1383 		cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1384 #endif
1385 		taskq_destroy(system_taskq);
1386 		return (1);
1387 	}
1388 
1389 	dynamic_taskq = taskq_create("spl_dynamic_taskq", 1,
1390 	    maxclsyspri, boot_ncpus, INT_MAX, TASKQ_PREPOPULATE);
1391 	if (dynamic_taskq == NULL) {
1392 #ifdef HAVE_CPU_HOTPLUG
1393 		cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1394 #endif
1395 		taskq_destroy(system_taskq);
1396 		taskq_destroy(system_delay_taskq);
1397 		return (1);
1398 	}
1399 
1400 	/*
1401 	 * This is used to annotate tq_lock, so
1402 	 *   taskq_dispatch -> taskq_thread_spawn -> taskq_dispatch
1403 	 * does not trigger a lockdep warning re: possible recursive locking
1404 	 */
1405 	dynamic_taskq->tq_lock_class = TQ_LOCK_DYNAMIC;
1406 
1407 	return (0);
1408 }
1409 
1410 void
1411 spl_taskq_fini(void)
1412 {
1413 	taskq_destroy(dynamic_taskq);
1414 	dynamic_taskq = NULL;
1415 
1416 	taskq_destroy(system_delay_taskq);
1417 	system_delay_taskq = NULL;
1418 
1419 	taskq_destroy(system_taskq);
1420 	system_taskq = NULL;
1421 
1422 	tsd_destroy(&taskq_tsd);
1423 
1424 #ifdef HAVE_CPU_HOTPLUG
1425 	cpuhp_remove_multi_state(spl_taskq_cpuhp_state);
1426 	spl_taskq_cpuhp_state = 0;
1427 #endif
1428 }
1429